CoRL 2026Poster

RoMo-GSEfficient Online 3D Gaussian Splatting
for Robot Mapping via
Exposure-Integrated Motion Blur

Sangwoo Jung1,2Di Deng1Jiahui Fu1Laura Herlant1*Ayoung Kim2*Yafei Hu1*

1 RAI Institute, United States2 Seoul National University, South Korea

* Co-corresponding authors

Paper Video Poster BibTeX
Motion-blurred observations from a handheld rig, a Spot robot, and a Franka arm are mapped into a compact 3D Gaussian scene with sharp rendered views.
From motion-blurred observations to sharp 3D maps. RoMo-GS reconstructs scenes online from a visual tracker's output, jointly refining the camera state and Gaussian map.
O(1)Exposure integration per Gaussian
15–150×Faster than offline deblurring baselines
~20%Fewer Gaussians than MODP

Abstract

Robot mapping,
with motion in mind.

Sharp renderings. Compact maps.
Online reconstruction.

Online 3D Gaussian Splatting (3DGS) is emerging as a practical representation for robotics. Real-world deployment, however, brings two challenges: rapid robot motion produces blurry observations that corrupt geometry and camera poses, while continuous mapping can grow the number of Gaussians without bound.

RoMo-GS addresses both. Instead of averaging multiple discrete sub-frames, we analytically integrate each projected Gaussian over the exposure interval, yielding O(1) complexity per Gaussian. A Hessian-based pruning score controls map growth, while joint Adam optimization refines poses, velocities, and Gaussians, followed by an online Levenberg–Marquardt step on the Gaussians.

Across public and self-collected datasets, RoMo-GS achieves competitive rendering quality with offline deblurring methods at substantially lower running time, and retains about 20% fewer Gaussians than the online MODP baseline at comparable quality.

Watch the system

RoMo-GS in action.

03:00 · Project video

Method

One exposure. One rasterization pass.

RoMo-GS combines exposure-integrated motion blur, joint optimization, and online density control in a tracker-agnostic mapping backend.

RoMo-GS pipeline: visual tracking, Gaussian insertion, exposure-integrated motion blur, blur-aware AccuTile, joint Adam updates, online LM refinement, and Hessian-based pruning.
Landmarks and depth samples initialize Gaussians. The backend models motion blur in a single pass and produces a compact map with sharp online renderings.
01

Exposure-integrated
motion blur

Analytically evaluate each projected Gaussian's exposure-time integral using the error function, under a linear screen-space motion approximation.

02

Single-sample
blur rasterization

Render a motion-blurred frame in one CUDA rasterization pass. The per-Gaussian cost stays constant instead of growing with the number of temporal samples.

03

Online optimization
& density control

Jointly refine pose, velocity, and Gaussians with Adam, apply an online LM step to the map, and remove redundant Gaussians with a Hessian-based score.

Motion-blurred ETH3D observation of tabletop objects.Sharper RoMo-GS reconstruction of the same tabletop scene.RoMo-GSBlurry input

Citation

BibTeX

@article{jung2026romo,
  title={RoMo-GS: Efficient Online 3D Gaussian Splatting for Robot Mapping via Exposure-Integrated Motion Blur},
  author={Jung, Sangwoo and Deng, Di and Fu, Jiahui and Herlant, Laura and Kim, Ayoung and Hu, Yafei},
  journal={arXiv preprint arXiv:2610.24605},
  year={2026}
}

Acknowledgment

We thank ETH Zürich for providing experimental space for real-world testing, data collection, and demo video filming.

Seoul National UniversityDepartment of Mechanical Engineering